Eric Horvitz Proposes Decision-Analytic Steering for High-Stakes LM Decisions
erichorvitz · x · 2026-08-04
Microsoft Chief Scientific Officer Eric Horvitz explores linking large models with decision theory under uncertainty to support consequential actions, such as emergency triage.
He notes that when AI systems translate uncertain predictions into actions, different types of errors often have asymmetric consequences. Therefore, system design must move beyond simple model predictions to consider the decision policies that convert predictions into actions and reflect specific priorities.
More from AGI Musings
- Introducing Pax Machina: A new magazine on AI-era institutions — xuanalogue · 2026-08-05
- Future of Life Institute Warns: AI Could Boost Economy Yet Bankrupt Governments — cccalum · 2026-08-05
- Introducing Pax Machina: New Publication on Institutions for Powerful AI — edelwax · 2026-08-04
- How AI Writing Tools Are Quietly Ruining Client Taste — JealousQuality3052 · 2026-08-04
- Why Experts Still Outperform LLMs: The Value of High-Quality Supervision — tokenbender · 2026-08-04
- How Mathematicians Cope with AI: Some Pay for ChatGPT Over Free Institutional Gemini — joshua_saxe · 2026-08-04